claude-code - 💡(How to fix) Fix Subagent ignores tool-call denials and keeps fetching instead of summarizing existing context [1 pull requests]

Summary

When a user denies a tool call inside a running subagent (e.g. denying a WebFetch), the agent should treat this as a signal to stop gathering and summarize what it already has. Instead, it treats each denial as a temporary obstacle and immediately tries the next URL on its list — continuing to accumulate tool uses and cost.

Steps to reproduce

1. Spawn a subagent with a research prompt (e.g. "search Reddit and HN for user opinions on X")
2. The agent starts fetching URLs in sequence
3. User denies one of the WebFetch tool calls via the permission prompt
4. Expected: agent stops fetching, uses its existing context to produce the summary
5. Actual: agent finds the next URL and tries again — indefinitely

Impact

• In this specific case: ~98 WebFetch calls were made, costing ~$1.15, before the session was force-killed via Ctrl+C
• The user explicitly tried to stop the agent mid-run by denying tool calls — the signal was ignored
• The useful context from the first ~98 fetches was lost because the agent never got to summarize

Expected behavior

A tool-call denial inside a subagent should be treated as "stop and use what you have." The agent has already gathered substantial context — it should synthesize that into a response rather than continuing to search.

This is similar to how a human researcher would respond to "stop, tell me what you found so far" — not by trying to open another browser tab.

Suggested fix

When a tool call is denied inside a subagent:

• Inject a system message along the lines of: "The user denied this tool call. Stop gathering and summarize your findings from the context you already have."
• Or: treat repeated denials (2+) as an implicit stop_sequence

Environment

• Claude Code CLI
• Model: claude-sonnet-4-6 (subagent)
• Platform: macOS